Comment by clnq | Hacker News Reader

Comment by clnq | Hacker News Reader

I wish we could stop it from lecturing me about how unsafe what I do for work is.

I am a SWE and I work close to the metal. ChatGPT is helpful, but it always says I shouldn’t do what I’m doing and I should use safer abstractions for my work. I write the low level abstractions, that is my job. Even if I ask it not to, it soon (over a few dozen messages) forgets.

My friend is a doctor and it’s the same story. He is always being told to see a qualified doctor about his questions. It just comes off as patronizing.

Sam Altman said a while ago that he believes the future of LLMs involves the user being able to customize them. I hope he hasn’t changed his mind. I definitely want a model I can make just give me straight answers without the social responsibility/legal disclaimer attached to each.

Come to think of it… what happened to legal disclaimers buried deep in the EULA? Why are they burned into nearly every answer LLMs (including Claude and Llama2) give us?

If you click on the 3 dots next to your email in the bottom left, you can set "custom instructions" now. Basically, you can tell ChatGPT who you are, as well as how you'd like it to respond.

It's not perfect but it's coming along.

And of you use the playground or any of the third party ChatGPT frontends that use the API you can just set the system prompt accordingly (custom instructions is basically just the system prompt). It'll still apologize here and there, but if you tell it that it's now a low level systems engineer AI, it should stop lecturing you about the dangers of low level systems engineering.

Have you tried copilot? I have had success with gpt until recently it seems to have a short term memory and (more annoyingly) will seemingly hit the context window limit before it even responds, even for what I would guess are short answers, forcing me to start an entirely new conversation, making me look towards copilot or claude.

I definitely like that Copilot does as instructed without talking back or lecturing me.

I am encountering this exact same issue. My instructions are repeatedly ignored and it falls back to a "preferred" approach after a single useful response. I have practically memorized the disclaimers attached to the end of every response.

If you add custom instructions, could you try the same questions with and without them, and let us know how it goes? I’m just really curious.

Reply on news.ycombinator.com